27/05/2019

Motivation

Changing Distributions

IPCC (2012)

Bivariate Formulation

What about Dependence?

Question

The marginal distribution of temperature extremes is changing -


Is the dependence also changing?


If so by how much?


What does this mean for events like heatwaves?

Contents

  • Marginals

  • Dependence

  • Observations

  • Climate Models

Marginals

GEV Distribution

Let \(M\) hottest 3-day average temperature in a year: \[ \mathbb{P}(M < z) = \exp \left\{ - \left[ 1 + \xi \left(\dfrac{z-\mu}{\sigma}\right) \right]_+^{-1 / \xi} \right\},\] where \([v]_+ = \max \left\lbrace 0,v \right\rbrace\), \(\mu \in \mathbb{R}\), \(\sigma \in \mathbb{R}^+\) and \(\xi \in \mathbb{R}\).
(Fisher and Tippett 1928, Gnendenko 1943)

Transform our marginals

  • Address the trend in changing temperature
    First example: Model changing in location linearly
    Later: (Lazy) Assume stationarity in a given period

  • For dependence we want common marginals

  • Wlog we can remove the trend and transform the marignal distributions
    \(\mathbb{P}(M < z) = \exp(-1/z)\)



Now dependence!

Dependence

Bivariate Extreme Value Theory

Assume the distribution of \((M_{i}, M_{j})\) can be approximated with a bivariate extreme value theory distribution, then

\[\mathbb{P}\left( M_{i} \leq z_i,\, M_{j} \leq z_j \right) = \exp\left\{ - V_{ij} \left( \dfrac{-1}{\log F_i(z_i)}, \, \dfrac{-1}{\log F_j(z_j)}\right) \right\}\] where the exponent measure \(V_{ij}(a,b)\) is given by \[V_{ij}(a, b) = 2 \int_0^1 \max \left(\dfrac{w}{a}, \dfrac{1-w}{b} \right) \text{d}H_{ij}(w)\] and \(H\) is any distribution function on \([0,1]\) with expectation equal to 0.5.

Approach

Lots of parameteric options to specify form of \(V_{ij}(a, b)\), but

  • Want to avoid assuming that a parametric form that is constant in time

  • Want a solution that will work for lots of different pairs

so …



Nonparametric

Pickands Dependence Function

\[V_{ij}(z) = \left(\frac{1}{z_i} + \frac{1}{z_j}\right)A\left(\frac{z_i}{|z_i + z_j|} + \frac{z_j}{|z_i + z_j|}\right)\]

C1: \(A(t)\) is continuous and convex for \(t \in [0,1]\)
C2: Bounds on the function of \(\max \{(1 - t), t \} \leq A(t) \leq 1\)
C3: Boundary conditions of \(A(0) = A(1) = 1\)

Data Example: Logistic Copula

Fitted Example: Pickands Function

Estimators

Real stations

Summary statistic

\[\mathbb{P}\left(M_{i} \leq z, M_{j} \leq z\right) = \left[\mathbb{P}(M_{i} \leq z)\mathbb{P}(M_{j} \leq z) \right]^{A\left(1/2\right)}.\]

Climate Models

Background

Model information

  • 22 Global Climate Model - Regional Cliamte Model combinations from EURO-CORDEX

  • 0.11 degree simulation

  • rotated-pole grid

  • 1950-2005 with historical forcing

  • 2006-2100 with RCP 8.5

  • Consider Annual Maxima

Multi-model Mean of A(1/2)

Pairs relative to De Bilt, NL
Years modelled: 1950 - 2000

Changes in Spatial Extent

A(1/2) < 0.7

Density of A(1/2)

A(1/2) < 0.7

Conclusions

Conclusions

Strong initial evidence to suggest:



  • The dependence between temperature extremes is increasing



  • Area impacted by heat extremes is getting larger

Future Work

  • Examine assumptions on the marginal distriubtion

  • Fit dependence as a function of time

  • Think about how to translate insights from extreme value statistics into insights about changing heat extremes

  • That includes 'event' definition

–>